Comparative Analysis of Five Decomposition Techniques for Monthly Rainfall Time Series from 1991 to 2020 in Guinea (West Africa)

The selection of an optimal time series decomposition’ technique is crucial for enhancing rainfall prediction by using sophisticated methods as deep learning hybrid approaches. This study evaluates the effectiveness of five times series decomposition techniques in analyzing monthly rainfall datasets across Guinea geographical regions (Lower-Guinea (Boké, Conakry, Kindia), Middle-Guinea (Labé, Mamou), Upper-Guinea (Faranah, Kankan) and Forest-Guinea (Nzérékoré)) and identifies the most suitable technique for each region. The decomposition techniques considered are (a) Variational Mode Decomposition (VMD), (b) Ensemble Empirical Mode Decomposition (EEMD), (c) Complete Ensemble Empirical Mode Decomposition (CEEMD), (d) Complete Ensemble Empirical Mode Decomposition Adaptive Noise (CEEMDAN) and (e) Improved Complete EEMD with Adaptive Noise (ICEEMDAN) techniques. These techniques were chosen due to their widespread use and high accuracy. Their performance was judged using Kolmogorov-Smirnov Statistic (DKS), Root Mean Square Error (RMSE), Mean Square Error (MSE) and Mean Absolute Error (MAE). The results demonstrate that ICEEMDAN and CEEMDAN techniques showed the lowest RMSE, MAE, and MSE values compared to the others ones at all of the studied sites, indicating highest decomposition performances. However, ICEEMDAN consistently exhibited smaller DKS values compared to CEEMDAN, except in the site of Faranah and Nzérékoré. Therefore, ICEEMDAN is recommended as the most suitable technique for most studied stations, while CEEMDAN is preferable for Faranah and Nzérékoré. The analysis also revealed that monthly rainfall trends during the study period were predominantly nonlinear across all the sites. These nonlinear trends exhibited complex patterns such as alternating increases, suggesting significant climatic shift around the 2010s. The results highlight the limitations of traditional methods that assume linearity in rainfall trend detection, which may yield inaccurate conclusions. The finding provides valuable insights for improving rainfall forecasting in Guinea and enhancing our understanding of rainfall temporal variations in the period from 1991 to 2020.